How ChatGPT and Perplexity Pick Sources

ChatGPT and Perplexity pick sources by running a live web search for the query, re-ranking the results based on how directly each page answers the specific question, and selecting 2 or 3 that pass a quality bar. The bar is not backlink volume. It is direct-answer formatting, schema, freshness, and the domain's topical depth on the category you are in.

The rough pipeline both engines use

Both engines run a similar flow, though the re-ranking weights differ.

  1. Query rewriting. The user's question gets expanded into multiple search queries to widen coverage
  2. Web search. Real-time fetch of 20 to 50 candidate pages from Bing (ChatGPT Search) or Perplexity's own index
  3. Snippet extraction. Each candidate page gets its most relevant 1 to 3 sentences pulled out
  4. Re-ranking. The snippets are scored on relevance, clarity, freshness, and source credibility
  5. Answer generation. The top 2 or 3 snippets get stitched into a paragraph with the sources cited (searchengineland.com)

The whole pipeline runs in 3 to 10 seconds per answer. The re-ranking step is where pages are won or lost.

What the re-ranker rewards

The signals with the heaviest weight:

  • Direct-answer format. The question you are targeting gets answered in the first 2 to 3 sentences of the page or section. The re-ranker is trying to find quotable sentences, not read your whole essay
  • Schema markup. FAQ and HowTo schema declare that a page contains question-answer pairs. The re-ranker uses that as a shortcut to find the answer (schema.org)
  • Freshness. Pages updated in the last 3 to 12 months outrank pages with the same information from 2021 for most categories. AI answers are expected to be current
  • Topical depth. A domain with 30 pages on one narrow category outranks a domain with 3 pages on 10 unrelated categories. The re-ranker reads topical concentration as authority

Signals that matter less than you think:

  • Total backlink volume (still helps but not the lead)
  • Word count (3,000-word pages lose to 500-word pages that answer the question directly)
  • Keyword density in the traditional SEO sense

How ChatGPT differs from Perplexity

ChatGPT (in Search mode) leans on Bing's index. Pages that rank well on Bing tend to surface. Bing's ranker is more forgiving of thin backlinks than Google's, which gives newer sites a real shot (blogs.bing.com).

Perplexity runs its own index and its own re-ranker. Perplexity weighs recency and direct-answer density even more heavily than ChatGPT. A week-old page with a clean direct answer can outrank a 2-year-old page that is more established but more meandering.

Google AI Overviews (not strictly ChatGPT or Perplexity but often in the same conversation) pulls from Google's index with Google's ranking signals, meaning classical SEO still matters there more than for the other two.

What this means for your pages

Three practical rules, ranked by leverage.

  1. The first 2 sentences under every H1 and H2 must directly answer the question that heading implies. No throat-clearing, no "in today's fast-moving world" intros
  2. FAQ schema on every page that answers multiple related questions. HowTo schema on procedural pages. Article schema with author, datePublished, and dateModified on everything
  3. Refresh content quarterly. Update at least one sentence, update the dateModified, keep timestamps visible on the page

Do those three things and your citation share will improve without touching backlinks at all (otterly.ai).

The manual test

Before investing weeks into an AEO cluster, do this:

  1. Pick 10 questions your buyers actually ask
  2. Run each in ChatGPT Search AND Perplexity
  3. Screenshot the sources panel for each answer
  4. For each source, open the page and look at the first 2 sentences under the H1

You will see a pattern fast. The cited pages almost always have a clean direct answer in the opening lines. The pages that are not cited almost always have a 5-paragraph intro before they get to the point. That is the pattern to copy.